NBER WORKING PAPER SERIES
FOLLOW THE MONEY: METHODS FOR IDENTIFYING CONSUMPTION AND INVESTMENT RESPONSES TO A LIQUIDITY SHOCK
Dean Karlan Adam Osman Jonathan Zinman
Working Paper 19696 http://www.nber.org/papers/w19696
NATIONAL BUREAU OF ECONOMIC RESEARCH 1050 Massachusetts Avenue Cambridge, MA 02138 December 2013
The authors thank financial support from the Bill and Melinda Gates Foundation, the Consultative Group for Assistance to the Poor (CGAP) and AusAID. The authors thank Kareem Haggag, Romina Kazadjian, Megan McGuire, Faith McCollister, Mark Miller, and Sarah Oberst at Innovations for Poverty Action for project management and field support throughout the project, and the senior management and staff at First Macro Bank and FICO Bank for their support and collaboration throughout this project. The authors retained full intellectual freedom to report and interpret the results throughout the study. All errors and opinions are those of the authors. The views expressed herein are those of the authors and do not necessarily reflect the views of the National Bureau of Economic Research.
NBER working papers are circulated for discussion and comment purposes. They have not been peer- reviewed or been subject to the review by the NBER Board of Directors that accompanies official NBER publications.
© 2013 by Dean Karlan, Adam Osman, and Jonathan Zinman. All rights reserved. Short sections of text, not to exceed two paragraphs, may be quoted without explicit permission provided that full credit, including © notice, is given to the source. Follow the Money: Methods for Identifying Consumption and Investment Responses to a Liquidity Shock Dean Karlan, Adam Osman, and Jonathan Zinman NBER Working Paper No. 19696 December 2013 JEL No. D12,D92,G21,O12,O16
ABSTRACT
Identifying the impacts of liquidity shocks on spending decisions is difficult methodologically but important for theory, practice, and policy. Using seven different methods on microenterprise loan applicants, we find striking results. Borrowers report uses of loan proceeds strategically, and more generally their reporting depends on elicitation method. Borrowers also interpret loan use questions differently than the key counterfactual: spending that would not have occurred sans loan. We identify the counterfactual using random assignment of loan approvals and short-run follow-up elicitation of major household and business cash outflows, and estimate that about 100% of loan-financed spending is on business inventory.
Dean Karlan Jonathan Zinman Department of Economics Department of Economics Yale University Dartmouth College P.O. Box 208269 314 Rockefeller Hall New Haven, CT 06520-8629 Hanover, NH 03755 and NBER and NBER [email protected] [email protected]
Adam Osman Yale University 27 Hillhouse Avenue New Haven, CT 06511 [email protected] I. Introduction
What are the impacts of liquidity shocks on the consumption and investment decisions of households and small businesses? Answers to this question have implications for the theory, practice, and regulation of credit, as well as for modeling intertemporal consumer choice. They shed light on perceived returns to investment, and on the extent to which constraints bind more for some types of household spending than others. Estimating impacts of liquidity shocks matters in many domains, for example in understanding household leveraging and deleveraging decisions in the wake of credit supply shocks,1 as well as evaluating interventions such as business grants,2 unconditional cash transfers,3 and microcredit expansions.4
Papers that track responses to liquidity shocks often focus on estimating medium- and long- term effects by measuring spending patterns, balance sheets, or summary statistics of financial conditions several months or years post-shock. This reduced-form evidence has proven quite useful, but it often leaves the mechanism underlying any change unidentified. For each possible state of the world many months post-liquidity shock -- high enterprise growth relative to baseline, low enterprise growth, consumption growth, etc. -- there are many paths from the liquidity change to that outcome. Identifying mechanisms is important because different paths can have different welfare implications.
To take an example closest to the setting we examine in this paper, many microcredit impact evaluations do not find significant effects of microcredit on enterprise scale or profitability one or two years post-intervention, even when the loans are targeted to those who are microentrepreneurs at baseline.5 There are at least three possible explanations for these findings: 1) impacts only materialize over longer horizons due to compounded benefits, adjustment, etc. This hypothesis often motivates researchers and program advocates to highlight the value of longer-term outcome data; 2) microentrepreneurs do not actually invest marginal liquidity in their businesses, perhaps because they are credit constrained on the margin and have household investment or consumption smoothing with a higher expected return on investment (in utility terms) than business investment; 3) microentrepreneurs do invest microloan proceeds in their businesses, but these investments do not end up earning a positive net return.
The second and third explanations highlight the value of very short-run data on spending decisions post-shock: “following the money” from liquidity to spending decisions can reveal the
1 See e.g. Hall (2011), Eggertsson and Krugman (2012), and Mian and Sufi (2012). 2 See e.g. Fafchamps et al (2013), Karlan, Knight and Udry (2013), and de Mel, McKenzie and Woodruff (2008). 3 See e.g. Benhassine et al (2013), Blattman, Fiala and Martinez (2012), Haushofer and Shapiro (2013), Karlan et al. (2013). 4 See e.g. Angelucci, Karlan and Zinman (2013), Attanasio et al (2011) , Augsburg et al (2012), Banerjee et al (2013), Crepon et al (2011), Karlan and Zinman (2010), Karlan and Zinman (2011), and Tarozzi, Desai and Johnson (2013). 5 See the studies cited in the previous footnote, with the exception of Karlan and Zinman (2010), which examines untargeted consumer loans.
2 mechanisms underlying the paths from shock to outcomes. If the second explanation is accurate that motivates further attempts to identify causes, consequences, and cures for credit constraints. If the third explanations is accurate that motivates further attempts to understand why entrepreneurs make investments that, ex-post at least, do not yield a positive net return on average (Moskowitz and Vissing-Jorgensen 2002; Anagol, Etang, and Karlan 2013; Karlan, Knight, and Udry 2013).6
To take another example, Mian and Sufi (2011) find that borrowing against rising home values by existing homeowners drove a significant fraction of both the rise in U.S. household leverage from 2002 to 2006 and the increase in mortgage defaults from 2006 to 2008. How did homeowners deploy the borrowed funds? As the paper explains (p.2134):
The real effects of the home equity–based borrowing channel depend on what households do with the borrowed money. We find no evidence that borrowing in response to increased house prices is used to purchase new homes or investment properties. In fact, home equity–based borrowing is not used to pay down expensive credit card balances, even for households with a heavy dependence on credit card borrowing. Given the high cost of keeping credit card balances, this result suggests a high marginal private return to borrowed funds.
Knowing what sort of spending generates this high marginal private return would inform how economists specify consumer preferences, expectations, and other inputs into consumer choice models. For example, spend data would help distinguish liquidity constraints from self-control problems as drivers of leveraging, which Mian and Sufi highlight as a fruitful avenue for future research (p.2155).7
As both examples suggest, unpacking the mechanisms underlying the long-run effects of a liquidity shock may require data on consumption and investment choices immediately after the shock. If one can follow the money from liquidity shock to spending, it may help identify how households use liquidity to try to improve their lots.
But how exactly one might go about measuring spending in the immediate aftermath of a liquidity shock is not immediately obvious, methodologically speaking. There are several challenges. Administrative data is rarely available for the right sample, timeframe, or spending frequency, and even more rarely sufficiently comprehensive in its coverage of different types of consumption and investment. This makes survey design very important. Yet money is fungible, and household and (micro)enterprise balance sheets are often complex, so it may be cognitively difficult for survey respondents to identify the effects of the liquidity shock on their spending, relative to the counterfactual of no shock. Similarly, surveys that simply ask about past purchases
6 Now consider the opposite state of the world: say an evaluation of 12-month impacts does find that a microcredit expansion produces larger, more profitable businesses. The mechanism need not be investment in business assets per se (inventory, physical capital, etc.) Rather, it could be investments in human capital (training, health, child care, etc.) that enable the entrepreneur or business “helpers” from her family to be more productive. 7 For related inquiries see Bauer et al (2012), and Bhutta and Keys (2013).
3 produce noisy data, and measurement error increases with the length of the recall period (Nicola, Francesca, and Giné 2012). Moreover, surveys can produce biased rather than merely noisy data if respondents have justification bias, 8 worry about surveyors sharing information with tax authorities or a lender that “requires” loans be used for particular purpose, or feel stigma about using debt for consumption purposes (Karlan and Zinman 2008). In short, data constraints, strategic reporting, and respondent (mis)perceptions may all make it difficult to follow the money.
We address these challenges by comparing results from seven different methods for following the money obtained by borrowers subject to a randomized supply shock from one of two microlenders in Metro Manila or northern Luzon, Philippines. The majority of marginal borrowers (90%) —those close to the banks’ credit score cutoffs—were randomly assigned to be offered a loan, while the remaining potential borrowers (10%) were randomly rejected. As is typical in microlending, the loans are targeted to microentrepreneurial investment, and underwritten accordingly, but are not secured by collateral or restricted in their disbursement.
The first method for following the money uses three questions about intended loan uses on the banks’ loan applications. The second has a loan officer ask the borrower those same three questions shortly after loan disbursal. The third and fourth are nearly identical direct questions asked of borrowers, by independent surveyors, with no link to the bank, two weeks and two months after loan disbursal. The fifth is two “list randomization” questions, asked by independent surveyors two weeks after disbursal, that make it feasible for respondents to respond truthfully to sensitive questions without actually revealing details about their behavior (Karlan and Zinman 2012). The sixth and seventh use the lenders’ randomizations and the two-week and two-month independent follow-up surveys, by comparing a listing of recent expenditures (with no reference to recent borrowing) across the treatment and control groups.9
Before summarizing the results, we emphasize that our paper is more about demonstrating a methodological approach to identifying spending responses than about extrapolating substantive implications from our particular setting.
That said, the pattern of results suggests three key findings in our setting. First, respondents report strategically. They report very few non-business uses of loan proceeds to the bank, significantly more to independent surveyors when asked direct questions, and yet significantly more to independent surveyors when presented with lists of statements that allow them to report truthfully without directly revealing what they spent. Second, even when responding (more) truthfully, answers to questions about “did you spend X or more of your loan on…” are different
8 E.g., my business did not grow from last year to this year, so I won’t report (to the surveyor, or even perhaps to myself) that I actually did try to grow my business by investing in new assets earlier this year. 9 The randomization does actually produce a powerful “first-stage”: a substantial increase in borrowing for the treatment (loan approved) group relative to the control (loan rejected) group. This result is not surprising, given that Karlan and Zinman (2011) found a similar result with marginal borrowers from one of the same banks considered here.
4 than the counterfactual of greatest interest to economists and policymakers. For example, although 12% of our treatment group implicitly (via list randomization) reports spending 5,000 pesos (US$1 = 45 Philippine Pesos) or more of their most recent loan on a household expense in the independent survey two weeks post-randomization, the treatment group is no more likely than the control group to say yes to any of a long list of questions regarding household expenditures greater than 1,000 pesos during the past 2 weeks (the proportion is 13% in both groups, for an estimated treatment effect of zero). Third, we estimate that the treatment effect is actually entirely on business investment, specifically inventory. This treatment effect can account for the entire loan amount 2-weeks post-randomization, with even larger but more noisily estimated effects at 2-months post-randomization.
We believe the main methodological implication of our results is that researchers should consider collecting spending data on both treatment and control subjects very shortly after an exogenous liquidity shock. This data would complement longer-run follow-up data; e.g., in our setting, it will be interesting to see whether the short-run increases in inventory translate into long-run increases, and into higher profits.
II. Market Overview
We collected data with the cooperation of two different banks in the Philippines, one in Metro Manila (covering mostly peri-urban areas) and another in northern Luzon. Both banks are for-profit institutions that offer individual liability microloans at about 60% APR. Loan sizes range from 5,000 pesos to 50,000 pesos, with a mean (median) of 13,996 (10,000) in our sample. Loan maturities range from three to six months, with weekly repayments of principal and interest. Both banks require that applicants have an existing business, and be between 18 and 65 years old.
The Metro Manila bank has operated in the region since the 1960s. It had microloans outstanding to about 2,700 borrowers as of July 2013. This portfolio represents a small fraction of its overall lending, which also includes larger business and consumer loans, and home mortgages. Until the end of 2012 the bank’s microlending activities received subsidized technical assistance from a USAID-funded program.10
The second bank has operated in mostly rural areas of northern Luzon since the 1980s. It had microloans outstanding to 26,000 borrowers in 2011 and offers other financial products as well.
The microloan market in the Philippines is somewhat competitive, as described in Karlan and Zinman (2011). There are informal options as well, including moneylenders. For our purposes the key fact is that that rejected borrowers do not simply obtain credit elsewhere: our banks’ random assignments to credit actually do produce a substantial change in the total/net borrowing of applicants (see Section III-F below).
10The program was administered by Chemonics, Microenterprise Access to Banking Services (MABS).
5
Our sample is comprised of 1661 marginal loan applicants who were randomized into loan approval or denial (see Section III-B for details on the randomization). Table 1 Column 1 provides baseline descriptive statistics gleaned from loan applications. 81.7% of the sample are women, 73.5% are married, and 32.9% are college educated.11. The average applicant is 40.9 years old and has owned her business for 6.7 years. Nearly half of the businesses are “sari-sari” (corner/convenience) stores. 35.8% have regular employees/helpers (i.e., workers besides the owner), and average weekly cash flow in the businesses is 4,901 pesos (a bit more than $100).
III. Methods and Results
A. Overview To better understand how borrowers deploy loan proceeds, and report thereon, we follow individuals from when they first apply for a loan until two months later. By that endpoint, we suspect that most of any proceeds will have been spent; this seems like a reasonable assumption given the high interest rates and short maturities. Along the way we use a variety of different methods to try to get at the same underlying question: how did the loan change the client’s spending relative to a counterfactual in which the loan was not available?
Figure 1: Study Timeline
Loan First 2-3 Week 2 Month Application Repayment Survey Survey
Week 0123456789+ Number Loan Activity & Data Collection
Week 0: Week 1: Weeks 2&3: Week 8: Weeks 9-24:
Loan Activity: Loan Activity: Loan Activity: Loan Activity: Loan Activity: Applies for Loan First Repayment Continued Continued Continued Credit Score Calculated (For Treatment) Repayment Repayment Repayment Loan Randomization (For Treatment) (For Treatment) (For Treatment) Implemented Data Collection (Bank Employee): Data Collection Data Collection Data Collection Loan Use Questions (Independent (Independent (Bank Employee): Main Use of Funds Surveyor): Surveyor): Loan Use Questions Loan Use Questions Loan Use Questions Intended Use Listing List Randomization Spending Outflows Spending Outflows
11Females were not directly targeted by the bank. Enterprises of this size in the Philippines have greater female ownership; larger loans are serviced by a different part of the bank.
6
Our methods include various attempts to measure the counterfactual through direct elicitation (survey questions). They also include a method that combines less-direct elicitation of loan uses—by attempting to measure all recent large outflows from the household and business—with the random assignment of access to credit. The data come from four different interactions, with the same individual, over the course of about two months. Figure 1 summarizes the timeline and the data collected in four distinct steps: (1) an application for a bank loan by the individual; (2) a short survey of approved applicants at their first repayment, administered by a loan officer; (3) a questionnaire by an independent surveyor two to three weeks after the loan application and (4) a questionnaire by an independent surveyor about two months after the loan application.
B. Sample Creation and Randomization Our sample is comprised of 1,661 marginally creditworthy microloan applicants to the two banks described in Section II. Individuals applied from one of 16 bank branches at the Northern Luzon lender, or 8 branches at the Metro Manila lender, between July 2010 and March 2012. Each loan application is digitized by bank staff and credit-scored by underwriting software. For the purposes of this study, relatively small numbers of applicants with the highest (lowest) scores were automatically approved (rejected). The remaining applicants (about 85% of the pool) were randomly assigned to approval (with 90% probability) or rejection (with 10% probability).
This random allocation of loans to marginal clients serves as the identifying instrument for our analysis of the expenditure data described in Sections III-E and III-F below. Table 1 Column 2 confirms that the treatment and control groups are observably identical, in a statistical sense: regressing treatment assignment on treatment strata and the complete set of baseline characteristics in Table 1, we do not reject the hypothesis that the characteristics are jointly uncorrelated with treatment assignment (p-value = 0.488).
C. Data Collection Step 1: At Application, by Loan Officer The first pieces of data on loan uses come from loan applications. Applications are extensive, and take the form of loan officers interviewing applicants, reviewing their documents, and entering data into a small netbook computer. This process typically takes at least an hour to complete, and includes questions on income, household composition, assets and liabilities, and business cash flows.
The banks added three questions on loan uses to their applications at our behest. The applicant was first asked: (1) Do you plan to spend 5,000 pesos or more of your loan on any one household item? 12 (2) Do you plan to spend 2,500 pesos or more of your loan on servicing any other debt? Later the applicant was asked to provide a full listing of intended usage of the loan. The former two questions are designed to identify non-trivial non-business uses of loan proceeds, keeping in mind that the median loan size is 10,000 pesos and that borrowers may split loan proceeds among several different types of expenditures.
12 Exchange rate at time of surveys was US$1 = 43 Philippine Pesos.
7
This first step allows us to see how the applicants report their intended loan usage to the banks. These data will not be very informative about true intentions if applicants believe that their responses may affect the lender’s decision. For example, applicants might reasonably infer that banks prefer to lend exclusively for business purposes, and answer no to the questions about household and refinancing uses, regardless of their true intentions.
Table 2 Column 1 shows that very few applicants report non-businesses loan uses on their loan applications. Only 1.8% report planning to use their loan on a household transaction of 5,000 pesos or more (Panel A), and only 2.3% report planning to use their loan to pay down debt of 2,500 or more (Panel B). 13 Column 1 shows results for the treatment group only, for comparability with subsequent analysis. Results do not change if we include the control group.
Is the low reported prevalence of non-business uses on loan applications driven by strategic underreporting? Results below from steps 3 and 4 suggest yes, although only to a point. Before detailing those results we examine whether borrowers change their reporting behavior to the bank after they obtain a loan.
D. Data Collection Step 2: At First Loan Repayment, by Bank Credit Officer The second pieces of data on loan uses come from a very short survey, administered by loan officers to a subset of borrowers, at the time of first repayment (about one week after loan disbursal). The loan officers asked two questions designed to parallel the key questions from the application: (1) Did you spend, or do you plan to spend, 5,000 pesos or more of your loan on any one household item? (2) Did you spend, or do you plan to spend, 2,500 pesos or more of your loan on servicing any other debt?
This step allows us to check for differences between what applicants and borrowers tell the bank. We might see such differences if applicants misreported strategically in the first step and the main driver of that behavior was concern about getting approved for the first loan. On the other hand, several factors push against finding differences, including repeat contracting, and any desire among borrowers to appear consistent in their reporting behavior.
Table 2 Column 2 shows that reported prevalence of non-business uses post-loan is essentially unchanged from the loan application. Here we find less than one percent reporting using their loan on a large household transaction, while 2.9% report using it to pay down other debt.14 Sample size is lower in Column 2 because this step was implemented only at one bank and only for a short period of time. The data collection proved onerous for the bank, and the bank discontinued it after we observed the strong similarity in reporting behavior between this step (post-loan) and step one (application).
13 As we show in section 2 of the paper our randomization was successful and so comparing the reported loan use intentions of the treatment and control group will not be informative at this point. The only place where comparing the responses is useful is in columns 5 and 7, reported spending. 14The loan officers also asked the borrowers what they primarily spent their loans on and every borrower replied that they spent it on their business.
8
E. Data Collection Steps 3 and 4: 2-Week and 2-Month Surveys, by Independent Surveyor The third and fourth pieces of data on loan uses come from two surveys, administered by an independent surveyor about two weeks and two months after loan application, of both treatment and control group individuals. Surveyors located individuals at their place of business or home and invited them to take a survey on behalf of Innovations for Poverty Action (IPA), a research organization. Surveyors were not aware of any connection to the banks. Surveyors informed people in the sample frame that IPA obtained a list of potential survey respondents from a database of local businesses.15
Both surveys focus on direct elicitation of loan uses and the measurement of all recent substantial outflows, although the second survey is a bit shorter. Both were administered by the same surveyor. The scripts for key questions are reproduced in Appendix 1. Relative to the two- week survey, measuring outflows at two months has the potential advantage of allowing more time for all loan proceeds to be spent. It also has several potential disadvantages: more time for the control group to find alternative sources of financing (weakening power), a longer recall period (increasing measurement error), and/or more time for any short-run returns on investment to effect spending decisions (confounding inferences about the direct effect of borrowing on spending).
84% of our initial sample of 1,661 completed the first (two-week) survey. Table 1 Column 3 shows that treatment assignment does not significantly affect two-week survey completion. Column 4 shows, unsurprisingly, that baseline characteristics do predict survey completion. But Column 5 shows that these characteristics do not interact significantly with treatment assignment (p-value on the joint test = 0.239), offering reassurance that the treatment leaves the composition as well as proportion of survey respondents unchanged.
65.9% of our initial sample completed the second (two-month) survey. Table 1 Column 6 shows that treatment assignment does not significantly affect two-month survey completion. Column 7 shows, unsurprisingly, that baseline characteristics do predict survey completion. Column 8 shows that the interactions between baseline characteristics and treatment assignment are jointly significant; raising the possibility that treatment affects the composition of two-month survey respondents (Column 8) if not the response rate (Column 6).
The two-week survey begins with questions about basic demographics, health and savings. These introductory questions are designed to mitigate the likelihood that respondents infer any connection or association between the survey and their recent loan (application). The surveyor then asked the respondent for details on any outstanding loans, starting with the most recent one.
15 The goal was to be truthful yet also mask the relationship with the specific partnering bank. The surveyors themselves had no knowledge of the bank connection.
9
Respondents reporting a loan were then asked about their deployment of loan proceeds using three different methods.
First, the surveyor explicitly asked the two key loan use questions: (1) Did you spend 5,000 pesos or more of your loan on any one household item? (2) Did you spend 2,500 pesos or more of your loan on servicing any other debt? We expect the proportion of “yeses” here to be higher than those reported to the bank, since incentives for strategic misreporting to an independent surveyor should be lower. Table 2 Column 3 shows that this is indeed the case, to some extent. 5.5% of individuals report using a loan for a large household expense; compared to 1.8% on the loan application (the 3.7 percentage point difference has a p-value less than 0.001). 7.7% report using the loan to pay down other debt, compared to 2.3% on the loan application (the 5.4 percentage point difference has a p-value less than 0.001). Of course, borrowers may still underreport non-business uses if such uses are stigmatized, or if borrowers suspect a connection between the surveyor and their bank. Such concerns motivate our second elicitation method.
Second, the surveyor administered a list randomization exercise to elicit estimates of group- level proportions of respondents using loan proceeds to pay down debt or buy household goods. List randomization is used across various disciplines to mitigate the underreporting of socially or financially sensitive information (Karlan and Zinman 2012). The procedure asks a randomly- selected half of the respondents to report the total number of “yes” answers to four innocuous binary questions (Appendix 1), and the other half to report the total number of “yes” answers to the same four innocuous binary questions plus a fifth sensitive one. We did this separately for the two different loan use questions: (1) I spent over 5,000 pesos of my loan of a single household transaction” and (2) “I spent more than 2,500 pesos of my loan to pay down other debt.” We then estimate the proportion responding “yes” to the sensitive (loan use) question by subtracting the mean count of “yeses” for those who had only had the four innocuous questions from the mean count for those who had all five questions (including a loan use question).16 As expected, list randomization produces substantially higher estimates of non-business uses (Table 2 Column 4). We infer that 11.5% of respondents report spending at least 5,000 pesos of their loan proceeds on a single household transaction (p-value = 0.285), with 19.1% spending at least 2,500 of their loan proceeds on paying down other debt (p-value = 0.021).
All told, the results in Columns 1-4 suggest that elicitation method can have substantial effects on how borrowers report loan uses. Borrowers report more non-business uses when asked by an independent surveyor rather than a bank, and still more when they can report anonymously. The results suggest that list randomization, administered by an independent surveyor, produces relatively accurate estimates of how borrowers perceive their loan uses.
These results thus far do not address the question of how borrower perceptions accord with the reality that is most interesting to many researchers, practitioners, and funders: what did the
16 Those who do not report an outstanding loan instead are assigned the mean count of the short-list (innocuous, non-loan use questions only) group. Results are nearly identical if we instead drop these non-borrowers.
10 respondent buy that they would not have in the absence of the marginal loan? Fungibility may make it difficult to construct survey questions that elicit that counterfactual. For example, loan proceeds may be used to purchase inventory in the proximate sense of cash from bank being handed over to a supplier. But if the business owner would have purchased that inventory anyway, the marginal (counterfactual) purchase could be something else entirely; e.g., perhaps the cash flow that would have been used to purchase inventory is now used to purchase health care for an ailing family member.
The difficulty of identifying the counterfactual of interest motivates our third type of survey question: we ask each respondent to list each household and business outflow greater than 1,000 pesos from the past two weeks (type and amount).17 (Note the lack of any reference to loans or loan proceeds: this question asks about spending more broadly.) Together with the random assignment of loan approvals, we use responses to this question to identify the counterfactual: the impacts of the marginal loan on consumption and investment. Table 2 Column 5 reports the results, which show a striking lack of impact on non-business spending. The treatment (loan approved) and control (loan rejected) groups have identical proportions (0.133) of respondents reporting one or more household expenses >= 5000 pesos, for a treatment effect of zero (SE = 0.30). 18 For debt pay down, the treatment group has a slightly higher proportion (0.142 vs. 0.126), but the 1.6 percentage point difference is not statistically significant (p-value = 0.580).19
We find similar results, on a much higher base, in the two-month survey.20 Regarding the base, many more respondents directly report non-business uses, whether directly (Column 6) or on the outflow list (Column 7).21 Regarding the counterfactual of interest, when we compare the treatment group to the control group we find that the control group has an equally high base, statistically speaking. 22.7% of the treatment group report spending at least 5,000 pesos on any one household transaction while 18.0% of the control group does so. This difference of 4.7 percentage points is not statistically significant (p-value = 0.210). Similarly, 23.7% of the treatment group reports spending more than 2,500 pesos on other debt22 while 19.7% of the
17 Without any prompts for specific expense types. 18If we instead use a 1,000 peso cut-off we get an increase of 0.026 in treatment, (p-value=0.560). The cut-off at 5,000 pesos allows us to check for large household expenditures and lines up with the direct questions that are asked of the borrowers. 19 We are implicitly using the random assignment as an instrument for borrowing over the subsequent two weeks. The top rows of Table 3 confirms that the instrument is a powerful one; e.g., a treatment group member is 16 percentage points more likely to have a formal sector loan than a control group member. 20 The higher base could be due to respondents taking > 2 weeks to fully spend their loan proceeds, and/or to respondents’ increased comfort with the survey or surveyor. 21 We did not include list randomization on the two-month survey. 22 It may seem peculiar that the proportion of respondents who report spending more than 2,500 on debt pay down in the explicit question asked by the surveyor (column 6) is higher than the proportion that report this when listing out their spending over the past 2 months (column 7). This may be due to the fact that the outflow list has a 1,000 peso threshold, so if someone pays off debt in increments < 1,000 pesos but a total amount >= 2,500 pesos, the outflow list would miss this, whereas that direct question might capture it.
11 control group does so. This difference of 4.1 percentage points is not statistically significant (p- value = 0.291).
Taken together, the results in Table 2 highlight several key findings. Substantively, there is little evidence of substantial non-business uses of microenterprise loans in this particular setting. This is surprising, given low impact on business growth in general from microcredit (Angelucci, Karlan, and Zinman 2013; Attanasio et al. 2011; Augsburg et al. 2012; Banerjee et al. 2013; Crepon et al. 2011; Karlan and Zinman 2010), findings from a prior study with one of the lenders here that marginal borrowers decrease investment in their microenterprises (Karlan and Zinman 2011), and mounting concerns that people “over-borrow” to finance consumption (Zinman forthcoming).
Methodologically, we find that borrower reporting responds strongly to the elicitation method, and that direct elicitation of loan uses does not produce evidence on a key counterfactual—what borrowers purchase that they would not have purchased in the absence of a loan. Rather, we identify the counterfactual using random assignment of credit access coupled with short-term follow-up measurement of substantial outflows.
F. So Where Does the Money Go?
If the marginal expenditure financed by a loan is not on a household item or other debt service (Table 2), it presumably is on some sort of business investment. Can we actually detect an increase in business investment, or do measurement error or reporting biases make it futile to attempt to follow the money with survey data?
Tables 3 and 4 suggest that our methods can in fact identify the marginal spending: business inventory, in this case. We switch from the mean comparisons in Table 2 (Columns 5 and 7) to regressions to improve precision, and estimate OLS intention-to-treat (ITT) models, with Huber- White standard errors, of the form: